{"id":"edc56497-8494-488b-8952-dac507987442","arxiv_id":"2506.02733","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"LinkTo-Anime contributes 395 cel-anime-style clips with pixel-level forward and backward flow, occlusion masks, and line art, and shows fine-tuning on it improves optical flow and downstream anime production tasks.","lead":"A team built a new dataset of 2D anime-style videos with precise pixel-motion ground truth, generated by rendering 3D character models in cel-animation style. The dataset helps train optical flow algorithms for anime production tasks like in-betweening and colorization.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-production transfer is asserted from an in-distribution benchmark and a private, qualitative Cel comparison; the key generalization claim is not empirically pinned down.","rationale":"The reader's conditional verdict already identifies the same weak point: the synthetic pipeline's representativeness of real cel production is supported mainly by distribution similarity to a private Cel dataset. My stress-test sharpens this into a concrete evaluation concern. The numerical improvements in Table 2 are computed on the same rendering pipeline used for training, so they do not by themselves demonstrate transfer. The downstream evidence is also thin: Table 3a is on BasicPBC's own test set rather than real production frames, and Table 3b shows a very small PSNR/SSIM gain on Cel, with no uncertainty quantification. The missing RAFT+CreativeFlow+ baseline and the use of externally provided CreativeFlow+ weights further weaken the controlled comparison. This is not a fatal flaw: the concern is addressable by releasing a public validation subset and adding confidence intervals, and the dataset itself appears well-constructed with useful annotations. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":12501,"tokens_out":4081,"duration_ms":49791,"concrete_test":"Ask the authors to release a held-out subset of Cel (or collect a new public 500+ frame set from production files) with sparse manual point correspondences, then evaluate PWC-Net, GMA, GMFlow, and RAFT fine-tuned on LinkTo-Anime versus CreativeFlow+ and AnimeRun on that set, reporting EPE and PCK per motion-speed stratum with 95% bootstrap confidence intervals over clips. If T+LT does not beat T+Cr by more than the interval width, the claim that LinkTo-Anime improves real-production optical flow is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that LinkTo-Anime improves optical flow for cel-anime production rests on two evidential steps. First, Table 2 shows fine-tuning on LinkTo-Anime improves all EPE metrics, but this is measured on the LinkTo-Anime test set, generated by the same 3D rendering pipeline described in Secs. 3.1-3.2. A model fine-tuned on LinkTo-Anime is therefore tested in-distribution; these numbers measure fit to the rendering pipeline, not transfer to real production files. Second, the only quantitative evidence of production relevance is distribution similarity to a private 1,100-frame Cel set (PCA, t-SNE, luminance KL-divergence, Sec. 3.3) plus qualitative flow visualizations and a downstream interpolation gain of +0.47 PSNR and +0.002 SSIM on Cel (Table 3b). No error bars or significance tests are reported, the Cel set is not released, and the manual keypoints used for Cel evaluation are described only as 'coarse ground truth' in Sec. 4.2. In addition, the CreativeFlow+ fine-tuned weights were borrowed from AnimeRun checkpoints rather than trained under the paper's own protocol (Sec. 4.1), so the closest competing baseline is not controlled. The load-bearing premise—that the synthetic distribution transfers to real cel production—is therefore unverified; the claimed 'consistently improves' behavior could be an artifact of evaluating on the same distribution the models were fine-tuned on.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces LinkTo-Anime, a synthetic optical flow dataset for cel (celluloid) animation, generated by rendering 80 open-source VRoid models with Mixamo-driven motions. The dataset contains 395 video clips, roughly 29,000 frames at 1440x2560 resolution, split into train/validation/test, with forward and backward optical flow, occlusion masks, rendered color frames, and line-art versions. The authors benchmark PWC-Net, RAFT, GMA, and GMFlow under four fine-tuning protocols (Sintel, CreativeFlow+, AnimeRun, and LinkTo-Anime), report endpoint-error metrics on the LinkTo-Anime test set, and evaluate two downstream applications: line-drawing colorization with BasicPBC and frame interpolation with AnimeInterp. They also introduce a private 1,100-frame 'Cel' dataset of real production files used for distribution analysis and qualitative validation. The central claim is that LinkTo-Anime is the first high-quality optical flow dataset tailored to cel animation production and that fine-tuning on it consistently improves optical flow prediction.","tokens_in":12790,"tokens_out":5154,"duration_ms":52379,"significance":"If the claims hold, LinkTo-Anime is a genuinely useful community resource: it is large-scale, high-resolution, and provides exact pixel-level flow labels and occlusion masks together with line-art renderings, which are scarce for the cel-animation domain. The benchmark across four optical flow methods and two downstream tasks is a solid starting point, and the public release on Hugging Face is a concrete strength. However, the strongest claims about generalization to real production files are not yet empirically pinned down, because the main quantitative evaluation is performed on the same rendering pipeline used to create the dataset, and the only cross-domain validation relies on a private, unreleased Cel set with sparse manual annotations and no error bars. The dataset contribution itself is valuable even if the production-transfer claim is later scoped down, but the paper as written overstates what the evidence supports.","major_comments":[{"comment":"The claim that fine-tuning on LinkTo-Anime 'consistently improves optical flow prediction performance for all EPE metrics' is demonstrated only on the LinkTo-Anime test set, which is generated by the same 3D rendering pipeline described in Secs. 3.1–3.2. This is an in-distribution evaluation: a model fine-tuned on LinkTo-Anime is tested on frames drawn from the same synthetic distribution. The abstract and conclusion should either scope this claim to the proposed benchmark or provide a cross-domain evaluation on an independently released real-production set.","section":"Abstract; §4.2, Table 2"},{"comment":"The evidence that LinkTo-Anime transfers to real cel production files relies on PCA/t-SNE visual overlap, luminance KL divergence, qualitative flow visualizations, and a private 1,100-frame Cel dataset that is not released. The downstream interpolation improvement on Cel (+0.47 PSNR, +0.002 SSIM) is reported without error bars or significance tests, and the manual keypoints used for evaluation are described only as 'coarse ground truth' in Sec. 4.2. To substantiate the production-generalization claim, the authors should release Cel (or a curated annotated subset) and report confidence intervals or paired significance tests.","section":"§3.3; §4.2; Table 3b"},{"comment":"The CreativeFlow+ baseline is not controlled under the paper's own training protocol. The authors state that because the CreativeFlow+ download link is unavailable, they used pretrained weights provided in AnimeRun, and they omit RAFT ft. T+Cr entirely. This makes the 'second-best' comparisons in Table 2 and the conclusions drawn from them in Sec. 4.2 unreliable: the Cr columns may come from models trained with different data schedules or iteration counts. The authors should retrain all baselines under their own protocol or clearly label the Cr results as inherited from a different setup.","section":"§4.1"},{"comment":"The benchmark reports a single run per model-dataset combination with no variance estimates. Given that the differences between the top methods can be as small as 0.10 EPE (GMA ft. T+LT versus PWC-Net ft. T+LT), the claimed ranking and the 'consistently improves' language are not robust without repeated runs or error bars. At minimum, the authors should report the number of runs and the observed variability for the key comparisons.","section":"Table 2"}],"minor_comments":[{"comment":"There are several typos: 'the very frist step', 'libe art output', 'Aadapting models' in Fig. 4, and 'his is likely' in Sec. 4.2. The manuscript needs a careful proofread.","section":"§3.1; §4.2; Fig. 4"},{"comment":"References [30] and [31] are the same paper (Deep Geometrized Cartoon Line Inbetweening); one should be removed.","section":"References"},{"comment":"The caption 'Qualitative result of two Animation Application' is inaccurate because both panels are quantitative tables; please change it to 'Quantitative results'.","section":"Table 3"},{"comment":"The caption reads '(a) Anime Interpolation (b) Anime Interpolation'; the first panel is line-drawing colorization and the second is frame interpolation, so the caption should distinguish them.","section":"Figure 8"},{"comment":"The citation 'Buter et al.' should be 'Butler et al.' for the MPI-Sintel reference.","section":"§2.2"},{"comment":"The sentence 'the flow directions ... are mostly concentrated around 0, 90, -90, and 180 degree;' has a typographical issue with the semicolon and 'degree'; please polish.","section":"§3.3"}],"recommendation":"major_revision","confidential_remarks":"The dataset construction and within-domain benchmark are solid enough to merit publication after revision, but the production-transfer claims are not yet supported. The private Cel set is a reproducibility concern: if the authors do not release it, they should either provide a public proxy or substantially weaken the claims in the abstract and introduction. The CreativeFlow+ baseline issue should also be fixed before the benchmark is used as evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"LinkTo-Anime is a real resource: 395 clips, ~30k frames at 1440×2560, with forward/backward flow, occlusion masks, line art, white backgrounds, and fixed camera per cut. That combination doesn't exist in prior datasets, and the release fills a genuine gap for cel-anime optical flow. The rendering pipeline (preprocessing VRoid models, Mixamo rigging, Cartoon Line Creator, shading) is described concretely enough to reproduce, and the authors benchmark four flow methods with consistent fine-tuning protocols.\n\nThe within-dataset benchmark is convincing: fine-tuning on LinkTo-Anime beats fine-tuning on Sintel, AnimeRun, and CreativeFlow+ on the LinkTo-Anime test set, with large absolute EPE reductions. The downstream checks on BasicPBC colorization and AnimeInterp interpolation are a plus, and the gains there (e.g., +1.66 Acc, +1.44 B-MIoU; +0.47 PSNR) are plausible.\n\nThe soft spot is the real-production generalization claim. The Cel set is private, only 1,100 frames, with sparse manual keypoints described as coarse ground truth, and the distribution-similarity evidence (PCA/t-SNE/luminance KL) is suggestive but not compelling. More importantly, Table 2 is entirely in-distribution: test frames come from the same rendering pipeline used for fine-tuning. That measures fit to the pipeline, not transfer to production files. The CreativeFlow+ baseline is also not controlled — the authors used AnimeRun-provided weights because the download link was dead — so the 'consistently improves' claim is less clean than it looks. These are addressable flaws, not fatal ones, and the authors are honest about the missing pieces (private Cel, coarse annotations, missing weights).\n\nFor the audience this paper serves — researchers in anime inbetweening, colorization, and domain-specific optical flow — the dataset itself is the contribution, and releasing it is enough to justify engagement. The paper would be stronger with a released Cel subset, error bars or significance tests, and a softer claim about production transfer.\n\nSend it to peer review. It deserves a serious referee, and the evaluation issues are fixable. My own verdict would be conditional acceptance, with the transfer claim trimmed.","headline":"A genuinely useful cel-anime optical flow dataset, but the real-production transfer claim is asserted rather than demonstrated.","tokens_in":13309,"tokens_out":2258,"would_cite":true,"duration_ms":22668,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Anime-specific flow dataset lifts accuracy across all tested models.","keywords":["optical flow","cel animation","anime dataset","3D rendering","line art","benchmark","frame interpolation","colorization"],"falsifier":"Collect a held-out set of production cel frames from several independent studios, obtain manual or semi-automatic optical flow annotations on a subset, and compare fine-tuning on LinkTo-Anime versus AnimeRun and CreativeFlow+ on those frames; if LinkTo-Anime no longer yields lower EPE or better downstream colorization/interpolation metrics, the central claim fails.","tokens_in":12349,"feed_emoji":"🎬","tokens_out":4750,"duration_ms":42067,"temperature":0.7,"pith_summary":"The paper introduces LinkTo-Anime, a dataset of 395 rendered video clips built to look like cel-anime production frames, with dense forward and backward optical flow, occlusion masks, line-art versions, and skeletal motion data. Its central claim is that this is the first high-quality optical flow dataset tailored to cel anime character motion, and that fine-tuning standard optical flow networks on it improves their endpoint error on every reported metric. If true, the dataset fills a gap left by existing synthetic and animation datasets, which either lack detail or do not match the white-background, stationary-camera style of real production cuts. A sympathetic reader would care because optical flow underpins tasks like anime line colorization and frame interpolation, and a domain-specific benchmark could make those tools more reliable.","feed_headline":"Anime flow dataset lifts accuracy across all tested models","feed_subtitle":"Fine-tuning PWC-Net, RAFT, GMA, and GMFlow on 30K rendered cel-style frames cuts endpoint error on every metric.","key_machinery":"The pipeline renders pre-processed 3D models with a cel-style appearance: Lambertian diffuse shading plus a normal map for light-dark contrast, and contour lines produced with the Cartoon Line Creator plugin refined by a weld modifier, avoiding the FLAT rendering used by AnimeRun to preserve fine details such as eye interior and collar structure. Ground-truth optical flow comes from Blender's vector motion, giving per-pixel forward and backward correspondence, and occlusions are derived from those maps. Fixed camera settings per clip match the reutilized-background convention of cel production.","core_discovery":"LinkTo-Anime provides 29,270 frames across 395 clips at 1440×2560 resolution, rendered from 80 open-source VRoid models that were manually refined, rigged with Mixamo skeletons, and animated with motions such as boxing and dancing, each clip shot from five fixed viewpoints that mimic typical cuts. The authors claim that fine-tuning PWC-Net, RAFT, GMA, and GMFlow on this dataset consistently improves EPE across background, foreground, occluded, and speed-stratified regions compared with fine-tuning on Sintel, CreativeFlow+, or AnimeRun, and that substituting the fine-tuned RAFT into BasicPBC colorization and AnimeInterp interpolation improves their metrics. The paper also demonstrates via PCA, t-SNE, and KL divergence of luminance histograms that LinkTo-Anime's distribution is closer to a private 1,100-frame 'Cel' set of real production files than the other public datasets.","pith_inferences":["The paper's distribution evidence is limited to a 1,100-frame private set; an independent evaluation against production files from other studios would settle whether the style match holds beyond that sample.","Because all clips come from 80 VRoid models with Mixamo motions, the dataset may underrepresent flat-color, low-detail or hand-drawn styles common in older cel animation; the benchmark's advantage might shrink on such imagery.","The rendering choices—Lambertian shading, normal-map details, and welded line art—set a new design point that future synthetic anime datasets could adopt, but the paper does not ablate each choice's contribution to the downstream gains.","The claimed supremacy might be specific to the four tested architectures; other flow networks or transformer-based models could respond differently to this domain shift."],"forward_implications":["Optical flow models fine-tuned on LinkTo-Anime generalize to the private Cel dataset, suggesting flow learned from these rendered frames transfers to real anime production files.","Substituting LinkTo-Anime-fine-tuned RAFT raises BasicPBC colorization accuracy to 82.94% and background MIoU to 60.82%, improving on fine-tuning with Sintel or AnimeRun.","AnimeInterp frame interpolation on Cel frames reaches higher PSNR (19.63) and SSIM (0.884) when its RAFT module is fine-tuned on LinkTo-Anime.","The dataset's white backgrounds and stationary camera setup make it a closer benchmark for production cuts than full-scene animation datasets, and models fine-tuned on it produce cleaner backgrounds and more accurate foregrounds."],"supporting_citations":[{"why":"Supplies the naturalistic 3D-rendered baseline dataset for optical flow evaluation and fine-tuning comparisons.","marker":"[1]"},{"why":"Provides the base optical flow training data mixed into every fine-tuning run at 1:10 ratio.","marker":"[17]"},{"why":"Is the prior 2D-animation flow dataset with clean backgrounds that LinkTo-Anime must beat on background regions.","marker":"[27]"},{"why":"Is the prior cel-style anime flow dataset whose FLAT rendering loses detail, the main point of comparison.","marker":"[29]"},{"why":"Is one of the four benchmarked flow networks and the source of RAFT weights used in downstream colorization and interpolation.","marker":"[35]"},{"why":"Is the benchmarked GMA architecture whose fine-tuning results are reported.","marker":"[12]"},{"why":"Is the benchmarked GMFlow architecture whose fine-tuning results are reported.","marker":"[40]"},{"why":"Is the benchmarked PWC-Net architecture whose fine-tuning results are reported.","marker":"[32]"},{"why":"Defines the BasicPBC line-drawing colorization task used to show downstream benefit.","marker":"[5]"},{"why":"Defines the AnimeInterp in-betweening task used to show downstream benefit.","marker":"[28]"}],"fun_headline_variants":["Anime flow dataset from 3D renders boosts four models' accuracy","30K rendered cel-style frames improve optical flow EPE","LinkTo-Anime: 3D-rendered flow data sharpens anime motion tracking","First cel-anime flow dataset cuts error across tested optical flow models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that LinkTo-Anime matches real cel production imagery rests on similarity to a single private 1,100-frame Cel set; if that set does not represent the broader anime industry, or if PCA/t-SNE and luminance KL divergence are insufficient evidence of style match, the dataset's real-world value is not established.","fun_headline_variants_meta":{"raw":{"variants":["Anime flow dataset from 3D renders boosts four models' accuracy","30K rendered cel-style frames improve optical flow EPE","LinkTo-Anime: 3D-rendered flow data sharpens anime motion tracking","First cel-anime flow dataset cuts error across tested optical flow models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000937,"raw_usage":{"total_tokens":3979,"prompt_tokens":886,"completion_tokens":3093,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":3014}},"tokens_in":502,"tokens_out":3093,"duration_ms":20199,"temperature":1.0,"reasoning_tokens":3014,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:17:01.728180+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect a held-out set of production cel frames from several independent studios, obtain manual or semi-automatic optical flow annotations on a subset, and compare fine-tuning on LinkTo-Anime versus AnimeRun and CreativeFlow+ on those frames; if LinkTo-Anime no longer yields lower EPE or better downstream colorization/interpolation metrics, the central claim fails.","supporting_citations":[{"cited_title":"A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation","cited_arxiv_id":null,"evidence_quote":"Provides the base optical flow training data mixed into every fine-tuning run at 1:10 ratio."},{"cited_title":"Creative flow+ dataset","cited_arxiv_id":null,"evidence_quote":"Is the prior 2D-animation flow dataset with clean backgrounds that LinkTo-Anime must beat on background regions."},{"cited_title":"Animerun: 2d animation visual corre- spondence from open source 3d movies, 2022","cited_arxiv_id":null,"evidence_quote":"Is the prior cel-style anime flow dataset whose FLAT rendering loses detail, the main point of comparison."},{"cited_title":"Raft: Recurrent all-pairs field transforms for optical flow, 2020","cited_arxiv_id":null,"evidence_quote":"Is one of the four benchmarked flow networks and the source of RAFT weights used in downstream colorization and interpolation."},{"cited_title":"Learning to estimate hidden motions with global motion aggregation","cited_arxiv_id":null,"evidence_quote":"Is the benchmarked GMA architecture whose fine-tuning results are reported."},{"cited_title":"Learning inclusion matching for anima- tion paint bucket colorization, 2024","cited_arxiv_id":null,"evidence_quote":"Defines the BasicPBC line-drawing colorization task used to show downstream benefit."},{"cited_title":"Metaxas, Chen Change Loy, and Ziwei Liu","cited_arxiv_id":null,"evidence_quote":"Defines the AnimeInterp in-betweening task used to show downstream benefit."}],"review_version":1}